summaryrefslogtreecommitdiff
path: root/mteval/ns_ter.cc
blob: 8c969e587458f3d9fc4562e994bdee7cd7f9fc0c (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
#include "ns_ter.h"

#include <cassert>
#include <iostream>
#include <limits>
#include <tr1/unordered_map>
#include <set>
#include <boost/functional/hash.hpp>
#include "tdict.h"

static const bool ter_use_average_ref_len = true;
static const int ter_short_circuit_long_sentences = -1;

static const unsigned kINSERTIONS = 0;
static const unsigned kDELETIONS = 1;
static const unsigned kSUBSTITUTIONS = 2;
static const unsigned kSHIFTS = 3;
static const unsigned kREF_WORDCOUNT = 4;
static const unsigned kDUMMY_LAST_ENTRY = 5;

using namespace std;
using namespace std::tr1;

namespace NewScorer {

struct COSTS {
  static const float substitution;
  static const float deletion;
  static const float insertion;
  static const float shift;
};
const float COSTS::substitution = 1.0f;
const float COSTS::deletion = 1.0f;
const float COSTS::insertion = 1.0f;
const float COSTS::shift = 1.0f;

static const int MAX_SHIFT_SIZE = 10;
static const int MAX_SHIFT_DIST = 50;

struct Shift {
  unsigned int d_;
  Shift() : d_() {}
  Shift(int b, int e, int m) : d_() {
    begin(b);
    end(e);
    moveto(m);
  }
  inline int begin() const {
    return d_ & 0x3ff;
  }
  inline int end() const {
    return (d_ >> 10) & 0x3ff;
  }
  inline int moveto() const {
    int m = (d_ >> 20) & 0x7ff;
    if (m > 1024) { m -= 1024; m *= -1; }
    return m;
  }
  inline void begin(int b) {
    d_ &= 0xfffffc00u;
    d_ |= (b & 0x3ff);
  }
  inline void end(int e) {
    d_ &= 0xfff003ffu;
    d_ |= (e & 0x3ff) << 10;
  }
  inline void moveto(int m) {
    bool neg = (m < 0);
    if (neg) { m *= -1; m += 1024; }
    d_ &= 0xfffff;
    d_ |= (m & 0x7ff) << 20;
  }
};

class TERScorerImpl {

 public:
  enum TransType { MATCH, SUBSTITUTION, INSERTION, DELETION };

  explicit TERScorerImpl(const vector<WordID>& ref) : ref_(ref) {
    for (unsigned i = 0; i < ref.size(); ++i)
      rwexists_.insert(ref[i]);
  }

  float Calculate(const vector<WordID>& hyp, int* subs, int* ins, int* dels, int* shifts) const {
    return CalculateAllShifts(hyp, subs, ins, dels, shifts);
  }

  inline int GetRefLength() const {
    return ref_.size();
  }

 private:
  const vector<WordID>& ref_;
  set<WordID> rwexists_;

  typedef unordered_map<vector<WordID>, set<int>, boost::hash<vector<WordID> > > NgramToIntsMap;
  mutable NgramToIntsMap nmap_;

  static float MinimumEditDistance(
      const vector<WordID>& hyp,
      const vector<WordID>& ref,
      vector<TransType>* path) {
    vector<vector<TransType> > bmat(hyp.size() + 1, vector<TransType>(ref.size() + 1, MATCH));
    vector<vector<float> > cmat(hyp.size() + 1, vector<float>(ref.size() + 1, 0));
    for (int i = 0; i <= hyp.size(); ++i)
      cmat[i][0] = i;
    for (int j = 0; j <= ref.size(); ++j)
      cmat[0][j] = j;
    for (int i = 1; i <= hyp.size(); ++i) {
      const WordID& hw = hyp[i-1];
      for (int j = 1; j <= ref.size(); ++j) {
        const WordID& rw = ref[j-1];
	float& cur_c = cmat[i][j];
	TransType& cur_b = bmat[i][j];

        if (rw == hw) {
          cur_c = cmat[i-1][j-1];
          cur_b = MATCH;
        } else {
          cur_c = cmat[i-1][j-1] + COSTS::substitution;
          cur_b = SUBSTITUTION;
        }
	float cwoi = cmat[i-1][j];
        if (cur_c > cwoi + COSTS::insertion) {
          cur_c = cwoi + COSTS::insertion;
          cur_b = INSERTION;
        }
        float cwod = cmat[i][j-1];
        if (cur_c > cwod + COSTS::deletion) {
          cur_c = cwod + COSTS::deletion;
          cur_b = DELETION;
        }
      }
    }

    // trace back along the best path and record the transition types
    path->clear();
    int i = hyp.size();
    int j = ref.size();
    while (i > 0 || j > 0) {
      if (j == 0) {
        --i;
        path->push_back(INSERTION);
      } else if (i == 0) {
        --j;
        path->push_back(DELETION);
      } else {
        TransType t = bmat[i][j];
        path->push_back(t);
        switch (t) {
          case SUBSTITUTION:
          case MATCH:
            --i; --j; break;
          case INSERTION:
            --i; break;
          case DELETION:
            --j; break;
        }
      }
    }
    reverse(path->begin(), path->end());
    return cmat[hyp.size()][ref.size()];
  }

  void BuildWordMatches(const vector<WordID>& hyp, NgramToIntsMap* nmap) const {
    nmap->clear();
    set<WordID> exists_both;
    for (int i = 0; i < hyp.size(); ++i)
      if (rwexists_.find(hyp[i]) != rwexists_.end())
        exists_both.insert(hyp[i]);
    for (int start=0; start<ref_.size(); ++start) {
      if (exists_both.find(ref_[start]) == exists_both.end()) continue;
      vector<WordID> cp;
      int mlen = min(MAX_SHIFT_SIZE, static_cast<int>(ref_.size() - start));
      for (int len=0; len<mlen; ++len) {
        if (len && exists_both.find(ref_[start + len]) == exists_both.end()) break;
        cp.push_back(ref_[start + len]);
	(*nmap)[cp].insert(start);
      }
    }
  }

  static void PerformShift(const vector<WordID>& in,
    int start, int end, int moveto, vector<WordID>* out) {
    // cerr << "ps: " << start << " " << end << " " << moveto << endl;
    out->clear();
    if (moveto == -1) {
      for (int i = start; i <= end; ++i)
       out->push_back(in[i]);
      for (int i = 0; i < start; ++i)
       out->push_back(in[i]);
      for (int i = end+1; i < in.size(); ++i)
       out->push_back(in[i]);
    } else if (moveto < start) {
      for (int i = 0; i <= moveto; ++i)
       out->push_back(in[i]);
      for (int i = start; i <= end; ++i)
       out->push_back(in[i]);
      for (int i = moveto+1; i < start; ++i)
       out->push_back(in[i]);
      for (int i = end+1; i < in.size(); ++i)
       out->push_back(in[i]);
    } else if (moveto > end) {
      for (int i = 0; i < start; ++i)
       out->push_back(in[i]);
      for (int i = end+1; i <= moveto; ++i)
       out->push_back(in[i]);
      for (int i = start; i <= end; ++i)
       out->push_back(in[i]);
      for (int i = moveto+1; i < in.size(); ++i)
       out->push_back(in[i]);
    } else {
      for (int i = 0; i < start; ++i)
       out->push_back(in[i]);
      for (int i = end+1; (i < in.size()) && (i <= end + (moveto - start)); ++i)
       out->push_back(in[i]);
      for (int i = start; i <= end; ++i)
       out->push_back(in[i]);
      for (int i = (end + (moveto - start))+1; i < in.size(); ++i)
       out->push_back(in[i]);
    }
    if (out->size() != in.size()) {
      cerr << "ps: " << start << " " << end << " " << moveto << endl;
      cerr << "in=" << TD::GetString(in) << endl;
      cerr << "out=" << TD::GetString(*out) << endl;
    }
    assert(out->size() == in.size());
    // cerr << "ps: " << TD::GetString(*out) << endl;
  }

  void GetAllPossibleShifts(const vector<WordID>& hyp,
      const vector<int>& ralign,
      const vector<bool>& herr,
      const vector<bool>& rerr,
      const int min_size,
      vector<vector<Shift> >* shifts) const {
    for (int start = 0; start < hyp.size(); ++start) {
      vector<WordID> cp(1, hyp[start]);
      NgramToIntsMap::iterator niter = nmap_.find(cp);
      if (niter == nmap_.end()) continue;
      bool ok = false;
      int moveto;
      for (set<int>::iterator i = niter->second.begin(); i != niter->second.end(); ++i) {
        moveto = *i;
        int rm = ralign[moveto];
        ok = (start != rm &&
              (rm - start) < MAX_SHIFT_DIST &&
              (start - rm - 1) < MAX_SHIFT_DIST);
        if (ok) break;
      }
      if (!ok) continue;
      cp.clear();
      for (int end = start + min_size - 1;
           ok && end < hyp.size() && end < (start + MAX_SHIFT_SIZE); ++end) {
        cp.push_back(hyp[end]);
	vector<Shift>& sshifts = (*shifts)[end - start];
        ok = false;
        NgramToIntsMap::iterator niter = nmap_.find(cp);
        if (niter == nmap_.end()) break;
        bool any_herr = false;
        for (int i = start; i <= end && !any_herr; ++i)
          any_herr = herr[i];
        if (!any_herr) {
          ok = true;
          continue;
        }
        for (set<int>::iterator mi = niter->second.begin();
             mi != niter->second.end(); ++mi) {
          int moveto = *mi;
	  int rm = ralign[moveto];
	  if (! ((rm != start) &&
	        ((rm < start) || (rm > end)) &&
		(rm - start <= MAX_SHIFT_DIST) &&
		((start - rm - 1) <= MAX_SHIFT_DIST))) continue;
          ok = true;
	  bool any_rerr = false;
	  for (int i = 0; (i <= end - start) && (!any_rerr); ++i)
            any_rerr = rerr[moveto+i];
	  if (!any_rerr) continue;
	  for (int roff = 0; roff <= (end - start); ++roff) {
	    int rmr = ralign[moveto+roff];
	    if ((start != rmr) && ((roff == 0) || (rmr != ralign[moveto])))
	      sshifts.push_back(Shift(start, end, moveto + roff));
	  }
        }
      }
    }
  }

  bool CalculateBestShift(const vector<WordID>& cur,
                          const vector<WordID>& hyp,
                          float curerr,
                          const vector<TransType>& path,
                          vector<WordID>* new_hyp,
                          float* newerr,
                          vector<TransType>* new_path) const {
    vector<bool> herr, rerr;
    vector<int> ralign;
    int hpos = -1;
    for (int i = 0; i < path.size(); ++i) {
      switch (path[i]) {
        case MATCH:
	  ++hpos;
	  herr.push_back(false);
	  rerr.push_back(false);
	  ralign.push_back(hpos);
          break;
        case SUBSTITUTION:
	  ++hpos;
	  herr.push_back(true);
	  rerr.push_back(true);
	  ralign.push_back(hpos);
          break;
        case INSERTION:
	  ++hpos;
	  herr.push_back(true);
          break;
	case DELETION:
	  rerr.push_back(true);
	  ralign.push_back(hpos);
          break;
      }
    }
#if 0
    cerr << "RALIGN: ";
    for (int i = 0; i < rerr.size(); ++i)
      cerr << ralign[i] << " ";
    cerr << endl;
    cerr << "RERR: ";
    for (int i = 0; i < rerr.size(); ++i)
      cerr << (bool)rerr[i] << " ";
    cerr << endl;
    cerr << "HERR: ";
    for (int i = 0; i < herr.size(); ++i)
      cerr << (bool)herr[i] << " ";
    cerr << endl;
#endif

    vector<vector<Shift> > shifts(MAX_SHIFT_SIZE + 1);
    GetAllPossibleShifts(cur, ralign, herr, rerr, 1, &shifts);
    float cur_best_shift_cost = 0;
    *newerr = curerr;
    vector<TransType> cur_best_path;
    vector<WordID> cur_best_hyp;

    bool res = false;
    for (int i = shifts.size() - 1; i >=0; --i) {
      float curfix = curerr - (cur_best_shift_cost + *newerr);
      float maxfix = 2.0f * (1 + i) - COSTS::shift;
      if ((curfix > maxfix) || ((cur_best_shift_cost == 0) && (curfix == maxfix))) break;
      for (int j = 0; j < shifts[i].size(); ++j) {
        const Shift& s = shifts[i][j];
	curfix = curerr - (cur_best_shift_cost + *newerr);
	maxfix = 2.0f * (1 + i) - COSTS::shift;  // TODO remove?
        if ((curfix > maxfix) || ((cur_best_shift_cost == 0) && (curfix == maxfix))) continue;
	vector<WordID> shifted(cur.size());
	PerformShift(cur, s.begin(), s.end(), ralign[s.moveto()], &shifted);
	vector<TransType> try_path;
	float try_cost = MinimumEditDistance(shifted, ref_, &try_path);
	float gain = (*newerr + cur_best_shift_cost) - (try_cost + COSTS::shift);
	if (gain > 0.0f || ((cur_best_shift_cost == 0.0f) && (gain == 0.0f))) {
	  *newerr = try_cost;
	  cur_best_shift_cost = COSTS::shift;
	  new_path->swap(try_path);
	  new_hyp->swap(shifted);
	  res = true;
	  // cerr << "Found better shift " << s.begin() << "..." << s.end() << " moveto " << s.moveto() << endl;
	}
      }
    }

    return res;
  }

  static void GetPathStats(const vector<TransType>& path, int* subs, int* ins, int* dels) {
    *subs = *ins = *dels = 0;
    for (int i = 0; i < path.size(); ++i) {
      switch (path[i]) {
        case SUBSTITUTION:
	  ++(*subs);
        case MATCH:
          break;
        case INSERTION:
          ++(*ins); break;
	case DELETION:
          ++(*dels); break;
      }
    }
  }

  float CalculateAllShifts(const vector<WordID>& hyp,
      int* subs, int* ins, int* dels, int* shifts) const {
    BuildWordMatches(hyp, &nmap_);
    vector<TransType> path;
    float med_cost = MinimumEditDistance(hyp, ref_, &path);
    float edits = 0;
    vector<WordID> cur = hyp;
    *shifts = 0;
    if (ter_short_circuit_long_sentences < 0 ||
        ref_.size() < ter_short_circuit_long_sentences) {
      while (true) {
        vector<WordID> new_hyp;
        vector<TransType> new_path;
        float new_med_cost;
        if (!CalculateBestShift(cur, hyp, med_cost, path, &new_hyp, &new_med_cost, &new_path))
          break;
        edits += COSTS::shift;
        ++(*shifts);
        med_cost = new_med_cost;
        path.swap(new_path);
        cur.swap(new_hyp);
      }
    }
    GetPathStats(path, subs, ins, dels);
    return med_cost + edits;
  }
};

#if 0
void TERScore::ScoreDetails(std::string* details) const {
  char buf[200];
  sprintf(buf, "TER = %.2f, %3d|%3d|%3d|%3d (len=%d)",
     ComputeScore() * 100.0f,
     stats[kINSERTIONS],
     stats[kDELETIONS],
     stats[kSUBSTITUTIONS],
     stats[kSHIFTS],
     stats[kREF_WORDCOUNT]);
  *details = buf;
}
#endif

} // namespace NewScorer

void TERMetric::ComputeSufficientStatistics(const vector<WordID>& hyp,
                                            const vector<vector<WordID> >& refs,
                                            SufficientStats* out) const {
  out->fields.resize(kDUMMY_LAST_ENTRY);
  float best_score = numeric_limits<float>::max();
  unsigned avg_len = 0;
  for (int i = 0; i < refs.size(); ++i)
    avg_len += refs[i].size();
  avg_len /= refs.size();

  for (int i = 0; i < refs.size(); ++i) {
    int subs, ins, dels, shifts;
    NewScorer::TERScorerImpl ter(refs[i]);
    float score = ter.Calculate(hyp, &subs, &ins, &dels, &shifts);
    // cerr << "Component TER cost: " << score << endl;
    if (score < best_score) {
      out->fields[kINSERTIONS] = ins;
      out->fields[kDELETIONS] = dels;
      out->fields[kSUBSTITUTIONS] = subs;
      out->fields[kSHIFTS] = shifts;
      if (ter_use_average_ref_len) {
        out->fields[kREF_WORDCOUNT] = avg_len;
      } else {
        out->fields[kREF_WORDCOUNT] = refs[i].size();
      }

      best_score = score;
    }
  }
}

unsigned TERMetric::SufficientStatisticsVectorSize() const {
  return kDUMMY_LAST_ENTRY;
}

float TERMetric::ComputeScore(const SufficientStats& stats) const {
  float edits = static_cast<float>(stats[kINSERTIONS] + stats[kDELETIONS] + stats[kSUBSTITUTIONS] + stats[kSHIFTS]);
  return edits / static_cast<float>(stats[kREF_WORDCOUNT]);
}